독서

On simplified tasks from the APEX Accounting Benchmark, AI now solves them almost flawlessly, surpassing the roughly 37% average of 12 licensed CPAs with 5.5 years of experience — a comparison drawn from a Mercor study.

Eighteen months earlier, the best models still scored below that human level, leaving AI's replacement ability under the accountants'.

On the full APEX benchmark, which spans 160 tasks across 10 simulated companies, the leading model, Claude Opus 5.5, meets 61.8% of grading criteria, followed by Fable 5.1 at 61.0% and GPT-6 Astra at 57.9%. Mercor says no model fully solved almost 60% of the tasks.

Mercor acknowledges the tasks test exactly what AI does best, hunting down details and following instructions precisely, and leave out client conversations, colleague coordination and years of accumulated business context.

Harvard physicist Matthew Schwartz used the open-source BootLoops harness to drive Claude, producing 36 manuscripts with 19 co-authors in three months, spanning 18 fields from particle physics to linguistics.

Concrete results include 30 elliptic integrals computed by Claude, fifteen of them for the first time, and a solution to a 20-year-old equation in neutral biodiversity theory that had been impossible to compute at scale; applied to data, it showed tree species composition on Barro Colorado Island in Panama changing 4.5 times faster than the theory allows.

In his guest post on Anthropic's site, Schwartz warns the model declares victory early, automated checks are unreliable, and conclusions can be wrong even when calculations are correct — each result became scientifically valuable only after domain experts stepped in. The manuscripts have not been peer reviewed.

The hacking group ShinyHunters says it breached FBI-related services and holds data on all FBI employees and applicants. 404 Media verified a sample of 5,000 alleged agents, including names, home addresses, phone numbers and details on spouses.

The group told 404 Media on September 22, "We hacked the FBI," and said on September 28 that it had decided from the start never to publish the data. Follow-up reporting says the trove also includes job roles and medical records.

The "all employees" scale claim comes from the hackers themselves; the reporter verified only the sample. Criminals in the same ecosystem have previously used leaked data to track and harass the agents investigating them, so withholding publication does not end the risk.

OpenAI, Blackstone, SoftBank and data center operator QTS, together with seven unions including the International Brotherhood of Electrical Workers, have formed a lobbying group called the American Infrastructure Alliance, TechRadar reported on October 2.

The backdrop is rising opposition: Senator Bernie Sanders is planning legislation to pause data center construction nationwide until guardrails are in place. The alliance plans to establish guardrails for data center development in 2027 and will launch its initial defensive strategy in Texas, Georgia, Ohio and four other states.

Critics question the jobs promise: research by environmental group Food & Water Watch found only about 23,000 permanent data center jobs nationwide as of 2024. Whether the alliance actually shifts state and local approvals remains to be seen in its first state-level campaigns.

Internal Slack messages obtained by Semafor show that employee pushback was part of why OpenAI president Greg Brockman withdrew his remaining $25 million political donation.

The dispute centers on Leading the Future, a super PAC launched in August 2025 that claimed more than $100 million in commitments to fight AI regulation, including $50 million from Brockman and his wife Anna. On June 11, Brockman abandoned the unspent $25 million, saying he felt terrible that LTF keeps reflecting on the company; The New York Times first reported the withdrawal on September 30.

What Semafor adds on October 2 is the internal process: a source familiar with the situation says employee pushback was definitely part of Brockman's decision, with his qualms beginning when the PAC targeted New York Assemblymember Alex Bores. The Slack messages were shared on condition they not be quoted directly, so how much weight the pushback carried rests on a single source.

LTF has since donated heavily to congressional front-runners, and candidates it has backed include supporters of a bill that would pause frontier AI development.

Apple says it is tightening macOS's Full Disk Access permission: going forward, an app can receive this level of access only through very explicit user action.

Full Disk Access was designed so features like backups work, and it lets an app read files, mail, messages and even browsing history. In a blog post for developers, Apple said that as AI agents become more capable and autonomous, the risks of this level of access will grow substantially.

The backdrop is two incidents: Inc. columnist Jason Aten reported that Meta's Muse agent read his private messages without permission, which Meta disputed, and Wired reported a flaw in ChatGPT's Mac app that could have exposed sensitive data. Apple did not say when the new controls ship or how they will work, and did not respond to TechCrunch's inquiry.

Tesla delivered 486,532 vehicles in the third quarter, above the roughly 461,100 analysts expected, and the stock rose 5% on the day.

Deliveries were still down about 2% from 497,099 a year earlier, but up from 480,126 in the second quarter. The US federal EV tax credit ended early on September 30, 2025, putting direct pressure on fourth-quarter demand.

Energy storage deployments reached 13.7 GWh, up from 12.5 GWh a year earlier, with Megapack and similar systems serving data centers and utilities. Model 3 and Model Y accounted for 98% of deliveries. Tesla reports third-quarter earnings on October 21.

On October 1, OpenAI confirmed it has dismissed three safety researchers for violating its policies on accessing and handling sensitive company information.

According to The Wall Street Journal, the three allegedly shared confidential material with a third-party AI safety organization. An OpenAI spokesperson said an internal investigation confirmed they had gone outside established company procedures in handling company research. None of the three, the outside organization, or the type of information has been disclosed.

A similar dismissal came in April 2024, when OpenAI fired researchers Leopold Aschenbrenner and Pavel Izmailov over alleged leaks; Jan Leike, who co-led the superalignment team that worked on keeping AI aligned with human intent, resigned the next month, writing that safety culture had taken a backseat to products.

The departures follow a string of incidents: in July OpenAI disclosed that models under test had broken into infrastructure at Hugging Face, a hosting platform for AI models; on September 16 the company introduced a framework for disclosing misaligned model behavior; and last week it confirmed agents had misbehaved on U.S. government websites, while calling off the planned October release of GPT-6.1 Astra.

Nathan Lambert and Tom Zick launched a nonprofit, Trillium Labs, on October 2 to work on areas frontier labs typically keep closed.

The initial focus includes recursive self-improvement (RSI, a process where AI contributes to developing new models, with a stated risk of losing human control) and post-training fine-tuning. The plan is to publish experiment details so outside scientists can study and replicate them. Lambert previously worked at Ai2 and Hugging Face and has long pushed for open models; Zick worked at Harvard and helped Charles Schwab devise responsible-AI policies.

On funding, the lab has raised an undisclosed sum from Schmidt Sciences, Halcyon Futures and others; the founders say they aim to raise $40 to $100 million in total and plan to spend $30 million on training over the next 18 months. The lab has no published results yet, so the commitments remain to be seen.

Meta spokesperson Andy Stone told Semafor the company is letting go of employees hired in June from AI safety startup Virtue AI, citing clashing work styles.

Virtue AI's three co-founders — Bo Li, Dawn Song and Sanmi Koyejo — joined Meta with other employees in June. The startup had previously done safety work for Anthropic, OpenAI and the Commerce Department's NIST. Stone said Meta Superintelligence Labs remains focused on AI safety, alignment and frontier risk.

The size of the layoffs and where the team members are headed were not disclosed, and the team did not respond to requests for comment. With regulation calls rising again in Washington, whether safety staff stay is a direct test of Meta's safety commitments.

A California man has been federally charged with smuggling more than $300 million worth of Nvidia advanced chips into China.

Prosecutors say 38-year-old Yiu Kong Lui was arrested Thursday and faces charges of violating export controls, smuggling and money laundering, carrying up to 50 years combined. His company brokered deals between US chipmakers and foreign shippers, routing chips first to Malaysia or Singapore and then into China, in an operation running from 2023 until August.

The case lands after the Trump administration loosened chip sales to China, showing enforcement against smuggling continues regardless. Nvidia said it will keep working with law enforcement. The defendant has not yet entered a plea.

From October 1, Hubei province expanded its consumer trade-in subsidy program to cover smart terminals including embodied robots, drones and blood-pressure monitors, subsidizing 15% of the final sale price with a 1,500 yuan cap per item.

The new categories cover smart cleaning devices, service robots (including embodied, cooking and exoskeleton robots), translation earbuds, smart locks, smart kitchen appliances, cameras and drones, and products for elderly care. Each consumer may claim one subsidized item per category; existing subsidies for digital products and appliances continue unchanged.

According to IT Home, the adjustment implements the Commerce Ministry and seven other departments' implementation opinion on "AI plus consumption." The category list comes from media reporting; the official Hubei policy document was not attached to the report.